Github user felixcheung commented on a diff in the pull request:
https://github.com/apache/spark/pull/18320#discussion_r123419616
--- Diff: R/pkg/inst/worker/daemon.R ---
@@ -30,8 +30,40 @@ port <- as.integer(Sys.getenv("SPARKR_WORKER_PORT"))
inputCon <- socketConnection(
port = port, open = "rb", blocking = TRUE, timeout = connectionTimeout)
+# Waits indefinitely for a socket connecion by default.
+selectTimeout <- NULL
+
while (TRUE) {
- ready <- socketSelect(list(inputCon))
+ ready <- socketSelect(list(inputCon), timeout = selectTimeout)
+
+ # Note that the children should be terminated in the parent. If each
child terminates
+ # itself, it appears that the resource is not released properly, that
causes an unexpected
+ # termination of this daemon due to, for example, running out of file
descriptors
+ # (see SPARK-21093). Therefore, the current implementation tries to
retrieve children
+ # that are exited (but not terminated) and then sends a kill signal to
terminate them properly
+ # in the parent.
+ #
+ # There are two paths that it attempts to send a signal to terminate the
children in the parent.
+ #
+ # 1. Every second if any socket connection is not available and if
there are child workers
+ # running.
+ # 2. Right after a socket connection is available.
+ #
+ # In other words, the parent attempts to send the signal to the children
every second if
+ # any worker is running or right before launching other worker children
from the following
+ # new socket connection.
+
+ # Only the process IDs of exited children are returned and the
termination is attempted below.
+ children <- parallel:::selectChildren(timeout = 0)
+ if (is.integer(children)) {
+ # If it is PIDs, there are workers exited but not terminated. Attempts
to terminate them
--- End diff --
right, I see your reference here
https://github.com/apache/spark/pull/18320#discussion_r122639738
but I'm not 100% getting it when looking at the source code
https://github.com/s-u/multicore/blob/e9d9bf21e6cf08e24cfe54d762379b4fa923765b/src/fork.c#L361
---
If your project is set up for it, you can reply to this email and have your
reply appear on GitHub as well. If your project does not have this feature
enabled and wishes so, or if the feature is enabled but not working, please
contact infrastructure at infrastruct...@apache.org or file a JIRA ticket
with INFRA.
---
---------------------------------------------------------------------
To unsubscribe, e-mail: reviews-unsubscr...@spark.apache.org
For additional commands, e-mail: reviews-h...@spark.apache.org
